SealGAN:基于生成式对抗网络的印章消除研究  

SealGAN:Research on the Seal Elimination Based on Generative Adversarial Network

作　　者：李新利[1] 邹昌铭 杨国田[1] 刘禾[1] LI Xin-Li;ZOU Chang-Ming;YANG Guo-Tian;LIU He(School of Control and Computer Engineering,North China Electric Power University,Beijing 102206)

机构地区：[1]华北电力大学控制与计算机工程学院,北京102206

出　　处：《自动化学报》2021年第11期2614-2622,共9页Acta Automatica Sinica

摘　　要：发票是财务系统的重要组成部分.随着计算机视觉和人工智能技术的发展,出现了各种发票自动识别系统,但是发票上的印章严重影响了识别准确率.本文提出了一种用于自动消除发票印章的SealGAN网络. SealGAN网络是基于生成式对抗网络CycleGAN的改进,采用两个独立的分类器来取代原本的判别网络,从而降低单个分类器的分类要求,提高分类器的学习性能,并且结合ResNet和Unet两种结构构建下采样-精炼-上采样的生成网络,生成更加清晰的发票图像.同时提出了基于风格评价和内容评价的综合评价指标对SealGAN网络进行性能评价.实验结果表明,与CycleGAN-ResNet和CycleGAN-Unet网络相比较, SealGAN网络不仅能实现自动消除印章,而且还能更加清晰地保留印章下的发票内容,网络性能评价指标较高.Invoice is an important part of the financial system.With the development of computer vision and artifi-cial intelligence technologies,various automatic invoice identification systems have been developed.However,the seals on the invoice often affect the identification success rate.In this paper,the SealGAN network structure is pro-posed,which can automatically eliminate the seal of invoice.SealGAN network is an improvement based on gener-ative adversarial network CycleGAN.The two discriminant networks are replaced with two independent classifiers in the SealGAN,which reduces the classification requirement of each classifier,and the learning performance of clas-sifiers can be improved.In addition,it combines the ResNet and UNet structures to construct a downsampling-re-fining-upsampling generation network,which can generate clearer images of invoice.And the network comprehens-ive evaluation index,including style evaluation and content evaluation,is proposed to evaluate the performance of network.Experimental results demonstrate that the SealGAN network can not only eliminate the seal automatic-ally,but also retain the invoice content under the seal clearly,and the network performance evaluation index is higher than that of the CycleGAN-ResNet and CycleGAN-Unet.

关 键 词：印章消除 生成式对抗网络 SealGAN CycleGAN 评价指标 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TP391.41[自动化与计算机技术—控制科学与工程] F275[经济管理—企业管理]